The creator-tooling lesson hiding in a Claude Code spending guard
Most social media operators I know are now running at least two AI subscriptions alongside their scheduling stack — one for copy and ideation, one for video or image generation — and almost none of them can tell you, mid-week, how much of that budget they’ve actually burned. That’s the operational gap that makes Peras worth ten minutes of your attention even though it has nothing to do with publishing. It’s a macOS utility from maker Kristiyan Ivanov that watches your Claude Code and Codex usage and refuses the next tool call once you cross a line you set, like 85% of your weekly allowance. For anyone whose content pipeline now leans on agentic AI — bulk caption generation, transcript cleanup, repurposing scripts — the interesting part isn’t the product. It’s the design pattern: a hard stop that fires during the run instead of a dashboard that reports the damage afterward.
What Peras actually solves, and why it’s a creator problem in disguise
Ivanov’s framing is refreshingly unglamorous. He says he kept running out of Claude Code mid-week and got tired of “constantly spamming /usage to guess how much a task would eat,” so he built a guard. You pick a threshold — his example is 85% of the week — and when usage crosses it, the next tool call is refused with a plain reason. You can resume for an hour or until the reset.
If you’ve ever watched an agent chew through a session trying to fix one stubborn test, you already understand the failure mode. In the launch thread, Anirudh Madhavan describes watching an agent “loop happily through a whole session’s budget on one stubborn test” — and Ivanov admits that exact scenario is why he built the thing. My own version of this: I’ve had a repurposing agent burn an afternoon re-editing the same hook variations because my prompt was underspecified. The output was fine. The cost was not.
Here’s the creator-economy angle, and it’s the reason I’m writing about a developer tool on a social-media blog. The way most of us use AI in content ops is bursty and agentic — you kick off a batch job and walk away. That’s precisely the usage pattern where a passive dashboard fails you. You don’t need a prettier chart at 6pm. You need the run to stop at 3pm.
The mechanic worth stealing: refusal with a reason
The clever bit, which commenter Fan flagged as “the part that stayed with me,” is that the agent is told why it stopped. Ivanov confirms the intent: “the agent gets told why it stopped, so you don’t have to keep checking.” That’s a meaningfully different design from a kill switch. A silent failure makes an agent retry or hallucinate workarounds; a explained refusal ends the turn cleanly.
If you build any internal automation — a Zapier-style chain, a custom script that pulls analytics and drafts posts — steal this. Fail loudly, fail with a human-readable reason, and fail before the expensive step rather than after.
How it differs from the tools you’re probably comparing it to
The honest answer is that Peras doesn’t have direct competitors in the social-media tool category, because it isn’t one. The right comparisons are the things it deliberately isn’t:
- Usage dashboards and spend trackers. The whole point, per Ivanov, is that the guard “stops the run itself, instead of reporting it after.” Reporting tools tell you what happened; this one changes what happens.
- Scheduling and publishing suites like Buffer, Later, or Metricool. These solve distribution and analytics for published content. Peras solves cost control for content production. They’re adjacent layers of the same workflow, not substitutes.
- General automation platforms like Zapier or Make. You could theoretically build a usage-check step in a workflow, but you’d be polling an API that, per Ivanov, doesn’t exist for this purpose — more on that below.
The pricing model is also worth noting because it’s the opposite of the SaaS norm creators are numb to. The CLI and hooks are free and MIT-licensed; the Mac app is €4.99 once. Ivanov’s own description: “basically paying me to sign and ship it,” and you can build it yourself from the repo. No subscription, no seat-based pricing, no “contact sales.” For indie creators running lean stacks, that’s a real signal about the maker’s incentives — he’s not optimizing for expansion revenue.
Why the data source matters more than the feature list
A commenter asked the obvious skeptical question: where does the weekly percentage actually come from? Ivanov’s answer is the most technically substantive part of the thread. Claude Code passes a rate_limits object to the status line containing five_hour and seven_day windows, each with used_percentage and resets_at. Peras reads rate_limits.seven_day.used_percentage and stores the latest reading. Crucially: “there’s no Anthropic usage API call and Peras isn’t estimating it from tokens.”
That distinction matters for trust. Token-estimation tools drift, because pricing tiers, caching, and model routing change the real cost of a token. Reading the platform’s own accounting number is a fundamentally more reliable approach — and it’s a pattern I’d bet we see more of as AI vendors expose usage telemetry to third-party tooling. If you’re evaluating any AI cost tool, ask this question first: is it reading the vendor’s number, or guessing?
What social media teams can borrow from this
Three transferable ideas, in descending order of how quickly you can implement them this week.
1. Set a per-task ceiling, not just a weekly one. The feature I find most useful for content ops is what Ivanov calls the per-prompt line: peras set prompt 3 pause means one prompt can’t burn more than 3% of your week no matter what the agent does inside it, and the next prompt starts fresh. Translate that to your own stack: cap the scope of any single AI batch job — one carousel, one week of captions, one video script — rather than trusting a global monthly budget to contain a runaway task. The global number is too coarse to catch a bad afternoon.
2. Use time-windowed lines for your publishing rhythm. Peras supports lines limited to hours of the day, e.g. peras set 09:00-18:00 weekly 80 pause. For a team that batches content Monday through Wednesday and publishes Thursday, a tighter line during production hours and a looser one during review days is a sensible shape. Ivanov notes per-project lines already work — run peras set --here weekly 80 pause in a repo and the line covers that folder and everything under it.
3. Treat your AI spend like ad spend. Growth marketers already know how to set daily caps on Meta Ads Manager and treat a blown budget as an incident. Very few apply the same discipline to the AI tools generating the creative that feeds those ads. That asymmetry is going to get expensive as agentic workflows become standard.
Why TikTok-first creators should care more than LinkedIn-first ones
This is my take, not the maker’s. Short-form video operations are the most iteration-heavy workflow in social: you’re generating dozens of hook variants, testing them, regenerating. That’s a high-volume, high-variance AI usage pattern where a runaway agent is most likely and most costly. A LinkedIn-first operator writing three posts a week from a single draft is running a much more predictable load. If your week looks like the former, a hard spending line isn’t a nice-to-have — it’s the difference between a bad Tuesday and a blown month.
Where my judgment says it falls short
The maker is unusually candid about limitations, and I’ll take him at his word rather than dress them up.
- macOS only. Linux “might work but isn’t tested,” and Linux testing is on his own to-do list. Windows isn’t mentioned at all. If your content team runs on Windows laptops — which plenty do — this isn’t for you yet.
- Subscriptions only. It’s built for subscription-based usage, not API-key billing. If you’re paying per token through the API for a custom pipeline, the
rate_limitsmechanism it depends on may not apply. - Codex hooks fail open. If the hook itself breaks, the guard doesn’t fire. Ivanov is explicit that this is a Codex limitation, “not something Peras can fix yet.”
claude -pcan’t be stopped. Non-interactive invocations escape the guard. For anyone running headless batch jobs — arguably the highest-risk usage pattern — that’s a real hole.- No loop detection. Ivanov is refreshingly blunt here: “5 edits to the same file is sometimes a loop and sometimes just TDD, so I’d rather cap the spend than guess.” I think that’s the right call, but it means the tool limits damage rather than preventing it.
- No day-of-week lines yet. Weekend vs. weekday thresholds aren’t supported. The workaround Ivanov offers is saving two profiles and switching manually on Friday with
peras profile use weekend— functional, but it’s a manual step that a distracted operator will forget.
Where the math breaks
A percentage threshold is only as good as the reset window it’s tied to. If a big batch job lands the day before your weekly reset, an 85% line might let a large task through that would have been refused on a Monday. There’s also a “pace line” that pauses when you’re on track to run out before reset, defaulting to 90% weekly — but the interaction between a pace line and a fixed threshold isn’t something I can evaluate from the thread, and I’d want to test it against my own usage curve before trusting it in production.
Who it’s not for: teams on Windows or Linux, API-billing shops, anyone whose AI usage is low-volume enough that a weekly glance at a dashboard suffices, and anyone who wants analytics rather than a guard. This is a narrow, sharp utility — and Ivanov seems to know it, describing the focus as “one sharp utility.”
What I’d watch / test next
The pattern here — read the vendor’s own usage number, refuse the next action with a reason, resume on your terms — is going to show up in more tools, and probably not just for coding agents. I’d watch for two things: whether Anthropic keeps exposing rate_limits to third parties (if it closes that door, this whole category gets harder), and whether the fail-open behavior on Codex hooks gets resolved upstream.
Concretely, this week: if you run Claude Code or Codex on a Mac, install the free CLI and set a line at whatever you’d consider an uncomfortable week — Ivanov is genuinely curious what percentage people choose, and the answer probably says something about how much slack your workflow needs. Then set a per-prompt cap on your most expensive recurring batch job and see whether it ever fires. If it does, you’ve found the task that was quietly eating your budget. If it doesn’t, you’ve learned your usage is more predictable than you feared — and you can raise the line with confidence. Either way, you’ll know more about your AI spend than the dashboard ever told you.
